[Bug]: Qwen-Image performance degrade when set tensor_parallel_size > 1 #1713
Description
Your current environment
The output of python collect_env.py
Your output of `python collect_env.py` here
Your code version
The commit id or version of vllm
0.16.0
The commit id or version of vllm-omni
90ec499bd557537088cbdf50259453b3a2f2d13e
🐛 Describe the bug
Reproduce code
from vllm_omni.entrypoints.omni import Omni
from vllm_omni.diffusion.data import DiffusionParallelConfig
from vllm_omni.inputs.data import OmniDiffusionSamplingParams
def main():
"""Generate an image of a cup of coffee on a table using the Omni model."""
omni = Omni(
model="Qwen/Qwen-Image",
vae_use_tiling=True,
parallel_config=DiffusionParallelConfig(
tensor_parallel_size=2,
vae_patch_parallel_size=2, # or 2
),
)
prompt = {
"prompt": "a cup of coffee on a table",
"negative_prompt": "low resolution",
}
sampling_params = OmniDiffusionSamplingParams(
num_inference_steps=10,
width=512,
height=512,
)
outputs = omni.generate(
prompt,
sampling_params_list=[sampling_params],
)
images = outputs[0].request_output[0].images
if images:
images[0].save("coffee.png")
print("✅ Image saved to coffee.png")
else:
print("⚠️ No images were generated.")
if __name__ == "__main__":
main()Here is result when run tensor_parallel_size=2
Here is result when run tensor_parallel_size=1
cc: @hsliuustc0106 @Gaohan123 @SamitHuang @ZJY0516
Before submitting a new issue...
- Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.